Abstract ID: 032
Assessing Decadal Variability of Subseasonal Predictability using Explainable Machine Learning
Lead Author: Marybeth Arcodia
Colorado State University, United States of America
Keywords: machine-learning, tropical variability, subseasonal, predictability
Abstract: The tropical MJO can influence midlatitude precipitation via tropical-extratropical teleconnections on S2S timescales, potentially resulting in life-threatening flooding and drought events. However, the field lacks an understanding of how the North American subseasonal precipitation predictability changes on low-frequency (~10-year) timescales. To assess low-frequency variability of predictability, we employ artificial neural networks to quantify how S2S prediction skill varies on decadal timescales along the West Coast of North America. The machine learning model uses tropical precipitation anomalies as a predictor of North American precipitation anomalies with a 3-week lead time to capture the subseasonal timescale of the tropical-extratropical teleconnection. Analysis of the prediction skill of the neural networks reveals fluctuations on decadal timescales, and the analysis is extended to observational data. The drivers of this low-frequency variability are investigated to understand why certain time periods have higher S2S predictability to identify forecasts of opportunity. Additionally, multiple eXplainable Artificial Intelligence (XAI) methods are used to identify sources of S2S precipitation predictability that are most strongly modulated on decadal timescales.
Co-authors:
Elizabeth Barnes (Colorado State University)
Kirsten Mayer (National Center for Atmospheric Research)
